library(nnet)
library(dplyr)## Warning: package 'dplyr' was built under R version 4.1.2##
## Attaching package: 'dplyr'## The following objects are masked from 'package:stats':
##
## filter, lag## The following objects are masked from 'package:base':
##
## intersect, setdiff, setequal, unionlibrary(tidyr)
set.seed(199)
rows <- sample(nrow(iris))
iris_rand <- iris[rows,]
train <- iris_rand[1:120,]
test <- iris_rand[121:150,]
nnet_fit <-nnet(Species ~ ., data = train, size=5, decay=5e-4, maxit=200)## # weights: 43
## initial value 160.443958
## iter 10 value 56.393902
## iter 20 value 53.611342
## iter 30 value 7.865334
## iter 40 value 7.022950
## iter 50 value 6.925476
## iter 60 value 6.891608
## iter 70 value 6.869529
## iter 80 value 6.850776
## iter 90 value 6.498715
## iter 100 value 5.020462
## iter 110 value 4.540204
## iter 120 value 4.427847
## iter 130 value 4.319533
## iter 140 value 4.163031
## iter 150 value 3.865988
## iter 160 value 2.120853
## iter 170 value 1.598539
## iter 180 value 1.485419
## iter 190 value 1.317687
## iter 200 value 1.275337
## final value 1.275337
## stopped after 200 iterationsAfter we have fitted our model we want to see how our model performs on new data so we will use the predict function to predict the species type of our test set
y_predL = predict(nnet_fit, newdata = test[,-5])
y_predL <- round(y_predL)
y_predL <- as.data.frame(y_predL)
y_predL <- y_predL %>% pivot_longer(everything()) %>% filter(value == 1) %>% mutate(name = as.factor(name))We can compare our predictions against the test set in a confusion matrix and explore the accuracy
Wow, nnet predicted the iris species with perfect accuracy.
caret::confusionMatrix(y_predL$name, test[,5])## Confusion Matrix and Statistics
##
## Reference
## Prediction setosa versicolor virginica
## setosa 10 0 0
## versicolor 0 11 0
## virginica 0 0 9
##
## Overall Statistics
##
## Accuracy : 1
## 95% CI : (0.8843, 1)
## No Information Rate : 0.3667
## P-Value [Acc > NIR] : 8.475e-14
##
## Kappa : 1
##
## Mcnemar's Test P-Value : NA
##
## Statistics by Class:
##
## Class: setosa Class: versicolor Class: virginica
## Sensitivity 1.0000 1.0000 1.0
## Specificity 1.0000 1.0000 1.0
## Pos Pred Value 1.0000 1.0000 1.0
## Neg Pred Value 1.0000 1.0000 1.0
## Prevalence 0.3333 0.3667 0.3
## Detection Rate 0.3333 0.3667 0.3
## Detection Prevalence 0.3333 0.3667 0.3
## Balanced Accuracy 1.0000 1.0000 1.0This package is very new to me. We are going to work through the installation and application as layed out in the tensorflow tutorial found here.
First we need to load the keras and tensorflow packages. You need to follow the tensorflow installation instructions otherwise you will run into trouble in short order.
That being said, I could not get ‘reticulate’ to build on my machine and ended up using ‘r-miniconda’ which I believe is why I’m am getting messages below.
library(keras)## Warning: package 'keras' was built under R version 4.1.2library(tensorflow)## Warning: package 'tensorflow' was built under R version 4.1.2We are using the mnist number dataset to perform some image classification. First we are loading the data into a training and test set. Then we are converting from integers to floating point.
c(c(x_train, y_train), c(x_test, y_test)) %<-% keras::dataset_mnist()
x_train <- x_train / 255
x_test <- x_test / 255Now we are building our model. We have an input layer. We flatten the data into a single vector. We create a densely connected nn layer. Our drop out layer extinguishes 20% of our signal (sets it to 0) to prevent overfitting
model <- keras_model_sequential(input_shape = c(28, 28)) %>%
layer_flatten() %>%
layer_dense(128, activation = "relu") %>%
layer_dropout(0.2) %>%
layer_dense(10)Our model returns log-odds for each class.
predictions <- predict(model, x_train[1:2, , ])
predictions## [,1] [,2] [,3] [,4] [,5] [,6]
## [1,] 0.35004312 0.4894799 0.16333592 0.09530532 -0.2847350 -0.8690768
## [2,] 0.06318349 0.4005950 -0.06793498 0.09019655 0.1042261 -0.1768343
## [,7] [,8] [,9] [,10]
## [1,] 0.5155104 0.01514847 0.7744809 -0.1699134
## [2,] 0.7210672 -0.27364829 0.7415282 -0.1057346Here we are converting log-odds to probabilities
tf$nn$softmax(predictions)## tf.Tensor(
## [[0.11629786 0.13369906 0.09649078 0.09014476 0.06164404 0.03436487
## 0.137225 0.08320105 0.17778811 0.06914447]
## [0.0862497 0.12086305 0.07565081 0.08861132 0.08986326 0.06784521
## 0.16652248 0.06158478 0.16996478 0.07284461]], shape=(2, 10), dtype=float64)Here we are defining our loss function that will be used in training.
loss_fn <- loss_sparse_categorical_crossentropy(from_logits = TRUE)“This loss is equal to the negative log probability of the true class” - TensorFlow for R
A loss of zero means we are certain of the class.
loss_fn(y_train[1:2], predictions)## tf.Tensor(2.9106146370132016, shape=(), dtype=float64)We need to configure and compile the model before it can be used.
model %>% compile(
optimizer = "adam",
loss = loss_fn,
metrics = "accuracy"
)This is pretty cool. Here we fit the model and get to see the decrease in loss and the improvement in accuracy.
model %>% fit(x_train, y_train, epochs = 5)Next lets check the accuracy. 97% not too shabby!
model %>% evaluate(x_test, y_test, verbose = 2)## loss accuracy
## 0.07202643 0.97770000Lastly we want to return the classes with the highest probability.
probability_model <- keras_model_sequential() %>%
model() %>%
layer_activation_softmax() %>%
layer_lambda(tf$argmax)probability_model(x_test[1:5, , ])## tf.Tensor([3 2 1 0 4 2 2 0 2 4], shape=(10), dtype=int64)